Agent have hard time recognizing MCP via executor

[krzyzanowskim]

I may use executor wrong, however the agents have a hard time recognize connected MCPs. When I instruct it to use certain MCP it says it has none. Only after I specify that MCP is available via executor MCP, the actor pick it up. Do you think it is solvable?

this is claude response:

[baggiiiie]

hey @krzyzanowskim , you can try to add search_tools=true to the mcp URL param , this exposes tools registered in executor to the harness directly

[StreamlinedStartup]

@baggiiiie I take it that also adds ALL tools for that person/workspace, as well as adds tool names (your screenshot) + tool descriptions i.e., token eat up?

[baggiiiie]

@StreamlinedStartup this depends on how your harness of choice implements its mcp client logic.

the problem @krzyzanowskim was facing was probably due to the harness deferring the mcp tool descriptions, most likely harness defers executor mcp description, while executor defers connection tools. therefor agent is having a hard time finding connections

harnesses and executor are doing this to avoid the exact token eating problem you described

two options are:

• have top level tools for each integrations, as shown in previous screenshot
• add a transparent mode where executor drops the codemode and exposes connection tools to the harness, this is what @RhysSullivan is doing in Add passthrough MCP mode #1942

hope this answers your question!